Method, device and equipment for determining the volume of three-dimensional point cloud of sand and gravel pile based on rail trolley

Abstract

The present application discloses a method, device, and equipment for determining the three-dimensional point cloud volume of a gravel pile based on a rail trolley, and relates to the technical field of bulk material measurement. The method comprises: pre-processing the original point cloud data of the gravel pile, and dividing the pre-processed point cloud data into regions using a curvature clustering segmentation algorithm; performing regional repair on the divided point cloud data using an adversarial network, and spatially dividing the repaired point cloud data based on an octree spatial topology index to obtain point cloud data at different levels, and determining the topological connection relationship between different levels to obtain a point cloud subset; using a convex hull-voxelization hybrid strategy to perform integral projection layered calculations, and then using a temporal point cloud training graph convolutional network to collaboratively optimize and calibrate the initial volume of the gravel pile to obtain a calibrated volume. The present application aims to quickly and accurately determine the three-dimensional point cloud volume of the gravel pile.

Description

Technical Field

[0001] The present application relates to the technical field of bulk material measurement, and in particular to a method, device and equipment for determining the three-dimensional point cloud volume of a gravel pile based on a rail trolley. Background Art

[0002] In fields such as construction, mining, logistics, and warehousing, accurate measurement of the volume of sand and gravel piles is directly related to material management, cost accounting, and project progress control. Traditional manual measurement methods rely on tools such as tape measures and total stations, which are subject to low efficiency, high safety risks, and significant subjective errors. In recent years, with the widespread adoption of 3D scanning technology, volume measurement methods based on LiDAR or photogrammetry have gradually replaced manual operations. However, measuring specialized bulk materials such as sand and gravel piles still faces significant technical challenges.

[0003] The surface of a gravel pile is formed by the random accumulation of irregular particles. This rough surface causes multiple reflections during laser scanning, resulting in a significant amount of outlier noise in the point cloud data. Research has shown that conventional median filtering or statistical filtering methods can mistakenly delete valid surface features (such as angular areas) when removing this multi-scale noise, resulting in the loss of critical geometric information. Furthermore, dust aerosols generated by field operations can obstruct the scanning line of sight, creating localized point cloud holes. Traditional Poisson reconstruction algorithms are prone to generating false surfaces when processing discontinuous missing areas, resulting in volume calculation errors exceeding 12%.

[0004] Sand and gravel piles contain a multi-level pore structure, including intergranular gaps (0.1cm-10cm) and interstitial gaps (10cm-50cm). Current voxelization methods typically use a fixed resolution to partition the space, making it difficult to strike a balance between computational efficiency and accuracy. While high-resolution voxels (e.g., 1cm³) can capture microscopic voids, they produce over a million voxel units, making real-time computation impossible. Low-resolution voxels (e.g., 10cm³) ignore over 50% of the effective void volume.

[0005] Current technological improvements are mainly focused on the following directions: (1) Using deep learning for point cloud completion, such as the PointNet++ network to repair missing areas, but it does not consider the continuity constraints of the normal vectors on the sand and gravel surfaces, and the repaired areas are prone to produce non-physically smooth surfaces; (2) Introducing a void ratio compensation coefficient, such as using an empirical formula (final volume V_real = volume V_scan × 1.15) for volume correction, but the fixed coefficient cannot adapt to the changes in sand and gravel with different particle sizes and humidity.

[0006] Therefore, the current field of gravel pile volume measurement still faces three major technical bottlenecks: (1) the lack of a robust point cloud processing framework for multi-scale noise and data loss; (2) existing void modeling methods do not establish a correlation model between local geometric features and global statistical laws; and (3) dynamic measurement systems have difficulty achieving an engineering balance between accuracy, efficiency, and cost. These shortcomings have led to the widespread use of existing equipment in industrial field applications, such as volume calculation errors exceeding 8% and high equipment failure rates, which has seriously restricted the development of intelligent management of bulk materials. Summary of the Invention

[0007] The purpose of this application is to provide a method, device and equipment for determining the three-dimensional point cloud volume of a sand and gravel pile based on a rail trolley, which can quickly and accurately determine the three-dimensional point cloud volume of a sand and gravel pile.

[0008] To achieve the above objectives, this application provides the following solutions:

[0009] In a first aspect, the present application provides a method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley, comprising:

[0010] Obtaining original point cloud data of the gravel pile; the original point cloud data is obtained by scanning and collecting based on the rail trolley;

[0011] Preprocessing the original point cloud data, and dividing the preprocessed point cloud data into regions using a curvature clustering segmentation algorithm to obtain divided point cloud data;

[0012] An adversarial network is used to perform regional repair on the divided point cloud data to obtain repaired point cloud data; the adversarial network is determined based on a normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field;

[0013] Performing spatial division on the repaired point cloud data based on an octree spatial topology index to obtain point cloud data of different levels, and determining topological connection relationships between different levels to obtain point cloud subsets;

[0014] A convex hull-voxelization hybrid strategy is used to perform integral projection layered calculation based on the point cloud subset to obtain the initial volume of the gravel pile.

[0015] A time-series point cloud training graph convolutional network is used to collaboratively optimize and calibrate the initial volume of the gravel pile to obtain a calibrated volume. The time-series point cloud training graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile. After performing internal compactness inversion, the time-series point cloud training graph convolutional network combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness-density-volume.

[0016] Optionally, the original point cloud data is preprocessed, and the preprocessed point cloud data is divided into regions using a curvature clustering segmentation algorithm to obtain divided point cloud data, specifically including:

[0017] Performing multi-scale noise filtering on the original point cloud data to obtain preprocessed point cloud data;

[0018] The curvature clustering segmentation algorithm is used to divide the preprocessed point cloud data into regions to obtain the divided point cloud data.

[0019] Optionally, a curvature clustering segmentation algorithm is used to divide the pre-processed point cloud data into regions to obtain divided point cloud data, including:

[0020] An improved Euclidean clustering method is used to calculate the point cloud curvature feature through principal component analysis; the point cloud curvature feature is determined based on the point cloud neighborhood feature;

[0021] The method for determining the point cloud curvature feature specifically includes:

[0022] For any target point in the preprocessed point cloud data, the local neighborhood is determined by K-nearest neighbor search and the covariance moment is constructed;

[0023] Based on the covariance moment, the eigenvalues ​​of the covariance matrix are obtained by eigenvalue decomposition; the eigenvalues ​​represent the distribution intensity of the neighborhood point cloud in three orthogonal directions;

[0024] Based on the eigenvalues, the formula is adopted:

[0025] ;

[0026] Determine the curvature characteristics of the point cloud;

[0027] For all target points in the preprocessed point cloud data, select the target point with the smallest value of the point cloud curvature feature as the seed point and create a new cluster ;

[0028] In the neighborhood of the seed point, filter and Join ;

[0029] right The neighborhood is expanded iteratively for the newly added points in and the search radius is adjusted dynamically:

[0030] ;

[0031] When there are no new points after a set number of consecutive iterations, cluster growth is completed;

[0032] Eliminate seed points when the number of points is less than Clustering, perform quadratic curvature check on the cluster boundary points to obtain the divided point cloud data;

[0033] in, For the The value of the point cloud curvature feature corresponding to the target point; is the first eigenvalue; is the second eigenvalue; is the third eigenvalue; For the The value of the point cloud curvature feature corresponding to the target point; To set the adaptive segmentation threshold; For the The target point and The Euclidean distance between target points; is the search radius; The lower limit of the number of seed points.

[0034] Optionally, the adversarial network includes: a discriminator and a generator; the discriminator includes an alternating structure of three-dimensional convolution kernel layers and maximum pooling layers; the generator includes a U-Net architecture with skip connections;

[0035] The loss function of the generator is determined based on the L1 regularization term and the normal vector direction cosine value using the normal vector constraint;

[0036] The expression of the loss function is:

[0037] ;

[0038] The expression corresponding to the normal vector constraint is:

[0039] ;

[0040] in, is the loss function; To rebuild losses; is the direction cosine value of the normal vector; To combat losses; is the normal vector constraint; is the expectation of the distribution of real point cloud data; It is the subnetwork in the discriminator that processes the normal vector field and is used to evaluate the authenticity of the normal vector field; is the normal vector field corresponding to the real point cloud; is the expectation of the noise distribution of the generator input; A fake point cloud generated by the generator based on noise; Fake point cloud The corresponding normal vector field.

[0041] Optionally, a convex hull-voxelization hybrid strategy is adopted to perform integral projection layered calculation based on the point cloud subset to obtain the initial volume of the gravel pile, specifically including:

[0042] Based on the Gaussian curvature of the local surface corresponding to the point cloud subset, according to Adjust the voxel size; where α is the voxel size coefficient; k1 is the local surface Gaussian curvature;

[0043] determining a voxel grid according to the adjusted voxel size;

[0044] Using a ray casting method, a projection layering process is performed based on the voxel grid to obtain a plurality of partitions;

[0045] According to the number of voxels in all partitions, the convex hull algorithm is used to perform integral accumulation processing to obtain the initial volume of the gravel pile.

[0046] Optionally, the temporal point cloud training graph convolutional network adopts a triple weight distribution mechanism to perform weighted fusion processing on the initial volume of the gravel pile to achieve collaborative optimization calibration and obtain a calibration volume; wherein the triple weight includes: global convex hull volume weight, local voxel volume weight and layered projection volume weight.

[0047] In a second aspect, the present application provides a device for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley, comprising:

[0048] A data acquisition module is used to obtain original point cloud data of the sand and gravel pile; the original point cloud data is obtained by scanning and collecting based on the rail trolley;

[0049] a processing module, configured to pre-process the original point cloud data and perform region division on the pre-processed point cloud data using a curvature clustering segmentation algorithm to obtain divided point cloud data;

[0050] A restoration module is configured to perform regional restoration on the divided point cloud data using an adversarial network to obtain restored point cloud data; the adversarial network is determined based on a normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field;

[0051] a partitioning module, configured to spatially partition the repaired point cloud data based on an octree spatial topology index to obtain point cloud data at different levels, and determine the topological connection relationship between the different levels to obtain point cloud subsets;

[0052] a calculation module, configured to adopt a convex hull-voxelization hybrid strategy to perform integral projection layered calculation based on the point cloud subset to obtain an initial volume of the gravel pile;

[0053] The calibration module is used to collaboratively optimize and calibrate the initial volume of the gravel pile using a time-series point cloud training graph convolutional network to obtain a calibrated volume; the time-series point cloud training graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile, performs internal compactness inversion, and then combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness-density-volume.

[0054] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for determining the three-dimensional point cloud volume of a gravel pile based on a rail trolley.

[0055] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0056] The present application provides a method, device and equipment for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley. The present application performs regional division based on a curvature clustering segmentation algorithm, and uses an adversarial network to perform regional repair on the divided point cloud data, which can extract the original point cloud data and achieve data integrity and accuracy. The octree spatial topology index can be used for efficient spatial division and hierarchical management, which can avoid redundancy in subsequent volume calculations. Then, the convex hull-voxelization hybrid strategy is used for integral projection hierarchical calculations, which can achieve coordinated processing of macro contours and micro gaps. Finally, the graph convolutional network is trained through time-series point clouds for collaborative optimization and calibration to improve the accuracy of volume determination. Therefore, the present application can quickly and accurately determine the volume of a three-dimensional point cloud of a gravel pile. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0058] Figure 1 This is a flow chart of a method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley;

[0059] Figure 2 Block diagram of the adversarial network;

[0060] Figure 3 Flowchart of the convex hull-voxelization hybrid strategy. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0062] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0063] In an exemplary embodiment, Figure 1 As shown, a method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley is provided, comprising:

[0064] Step 100: Obtaining original point cloud data of the gravel pile. The original point cloud data is obtained by scanning and collecting using a track-mounted trolley.

[0065] Step 200: pre-processing the original point cloud data, and using a curvature clustering segmentation algorithm to divide the pre-processed point cloud data into regions to obtain divided point cloud data.

[0066] Among them, the original point cloud data is preprocessed, and the curvature clustering segmentation algorithm is used to divide the preprocessed point cloud data into regions to obtain the divided point cloud data, which specifically includes:

[0067] Multi-scale noise filtering is performed on the original point cloud data to obtain preprocessed point cloud data.

[0068] The curvature clustering segmentation algorithm is used to divide the preprocessed point cloud data into regions to obtain the divided point cloud data.

[0069] In one embodiment, a curvature clustering segmentation algorithm is used to divide the pre-processed point cloud data into regions, and the divided point cloud data is obtained, including:

[0070] The improved Euclidean clustering method is used to calculate the point cloud curvature features through principal component analysis; the point cloud curvature features are determined based on the point cloud neighborhood features.

[0071] The method for determining the point cloud curvature feature specifically includes:

[0072] For any target point in the preprocessed point cloud data, the local neighborhood is determined by K-nearest neighbor search and the covariance moment is constructed.

[0073] Based on the covariance moments, the eigenvalues ​​of the covariance matrix are obtained through eigenvalue decomposition. The eigenvalues ​​represent the distribution intensity of the neighborhood point cloud in three orthogonal directions. The eigenvalues ​​include: the first eigenvalue, the second eigenvalue, and the third eigenvalue.

[0074] Based on the eigenvalue, the formula is used:

[0075] ;

[0076] Determine the point cloud curvature characteristics.

[0077] For all target points in the preprocessed point cloud data, select the target point with the smallest value of the point cloud curvature feature as the seed point and create a new cluster .

[0078] In the neighborhood of the seed point, filter and Join .

[0079] right The neighborhood is expanded iteratively for the newly added points in and the search radius is adjusted dynamically:

[0080] .

[0081] When there are no new points after a set number of consecutive iterations, cluster growth is completed.

[0082] Eliminate seed points when the number of points is less than The clustering is performed, and the quadratic curvature check is performed on the cluster boundary points to obtain the divided point cloud data.

[0083] in, For the The value of the point cloud curvature feature corresponding to the target point; is the first eigenvalue; is the second eigenvalue; is the third eigenvalue; For the The value of the point cloud curvature feature corresponding to the target point; To set the adaptive segmentation threshold; For the The target point and The Euclidean distance between target points; is the search radius; The lower limit of the number of seed points.

[0084] Step 300: Use an adversarial network to perform regional restoration on the divided point cloud data to obtain restored point cloud data. The adversarial network is determined based on the normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field.

[0085] The adversarial network includes: a discriminator and a generator; the discriminator includes an alternating structure of three-dimensional convolution kernel layers and maximum pooling layers; the generator includes a U-Net architecture with skip connections.

[0086] The loss function of the generator is determined based on the normal vector constraint and the L1 regularization term and the normal vector direction cosine value; the expression of the loss function is:

[0087] .

[0088] The expression corresponding to the normal vector constraint is:

[0089] .

[0090] in, is the loss function; To rebuild losses; is the direction cosine value of the normal vector; To combat losses; is the normal vector constraint; is the expectation of the distribution of real point cloud data; It is the subnetwork in the discriminator that processes the normal vector field and is used to evaluate the authenticity of the normal vector field; is the normal vector field corresponding to the real point cloud; is the expectation of the noise distribution of the generator input; A fake point cloud generated by the generator based on noise; Fake point cloud The corresponding normal vector field.

[0091] Step 400: spatially dividing the repaired point cloud data based on the octree spatial topology index to obtain point cloud data of different levels, and determining the topological connection relationship between different levels to obtain point cloud subsets.

[0092] Step 500: adopting a convex hull-voxelization hybrid strategy, performing integral projection layered calculation based on a point cloud subset to obtain the initial volume of the gravel pile.

[0093] In one embodiment, a convex hull-voxelization hybrid strategy is adopted to perform integral projection layered calculation based on a subset of the point cloud to obtain the initial volume of the gravel pile, specifically including:

[0094] Based on the Gaussian curvature of the local surface corresponding to the point cloud subset, according to Adjust the voxel size; where α is the coefficient of the voxel size; k1 is the local surface Gaussian curvature.

[0095] The voxel grid is determined according to the adjusted voxel size. The ray casting method is used to perform projection and layering processing based on the voxel grid to obtain multiple partitions. The convex hull algorithm is used to perform integral accumulation processing based on the number of voxels in all partitions to obtain the initial volume of the gravel pile.

[0096] Step 600: A graph convolutional network trained on a time-series point cloud is used to collaboratively optimize and calibrate the initial volume of the gravel pile to obtain a calibrated volume. This graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile. After performing internal compactness inversion, it combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness, density, and volume.

[0097] Then, the calibration volume is visualized in three dimensions to intuitively show the real-time volume of the gravel pile.

[0098] The convolutional network for training temporal point clouds adopts a triple weight distribution mechanism to perform weighted fusion processing on the initial volume of the gravel pile to achieve collaborative optimization calibration and obtain the calibrated volume. The triple weight includes: global convex hull volume weight, local voxel volume weight, and layered projection volume weight.

[0099] To address the challenges of rough surface, complex gaps, and dynamic deformation of gravel piles, this application proposes an innovative technical solution that integrates multimodal perception, deep learning restoration, and probabilistic error compensation. This technical solution includes the following steps:

[0100] Point cloud preprocessing: Multi-scale noise filtering is performed on the original point cloud scanned by the rail car. A normal-constrained generative adversarial network (GA-Net) is used to repair missing areas caused by dust occlusion and multiple laser reflections based on the surface characteristics of the sand and gravel. Combined with the curvature clustering segmentation algorithm, the point cloud is divided into geometrically consistent sand and gravel sub-regions, and a hierarchical structure is established through octree spatial topology indexing to optimize subsequent computational efficiency.

[0101] Dynamic volume calculation: A convex hull-voxelization hybrid strategy is used to construct a local coordinate system for each sub-region. The convex hull algorithm extracts the outer contour, and adaptive voxelization is used to capture the internal void distribution. Integral projection layered calculations are performed along the direction of sand and gravel accumulation, combined with a dedicated sand and gravel compensation model: local void ratios are quantified using X-ray backscatter data. A dynamic probabilistic error correction model is introduced to fuse multi-scale volume components (global convex hull volume, local voxel volume, and layered projection volume) for weighted optimization.

[0102] Specifically, the point cloud preprocessing uses multi-scale noise filtering to process the original point cloud data collected by the rail car, adopts a generative adversarial network with normal vector constraints to repair the missing areas, and uses a curvature clustering segmentation algorithm to divide the geometrically consistent sand and gravel sub-areas. (During the scanning process, the lidar scans not only sand and gravel, but also the silo environment. Therefore, it is necessary to divide the sand and gravel sub-areas, and then extract the point cloud data of the sand and gravel for subsequent volume calculations.) An octree spatial topology index is established. The octree spatial topology index is mainly used for efficient spatial division and hierarchical management of three-dimensional point cloud data of sand and gravel piles. Its core function is to construct a multi-resolution spatial structure through a non-uniform segmentation strategy, divide the point cloud into sub-areas of different levels, and establish a topological connection relationship between nodes. This hierarchical indexing significantly optimizes the computational efficiency of subsequent processing: During point cloud segmentation, a curvature clustering algorithm can rapidly locate geometrically consistent gravel sub-piles based on octree subregions, avoiding computational redundancy caused by global traversal. Furthermore, during dynamic volume calculations, a hybrid convex hull-voxelization strategy rapidly accesses point cloud subsets in the local coordinate system through octree indexing. Combined with adaptive voxel resolution adjustment, this strategy enables collaborative modeling of macroscopic contours and microscopic voids. Furthermore, bidirectional pointer connections ensure rapid querying of adjacent nodes, supporting the dynamic maintenance of spatial relationships during hierarchical integral projection calculations. By balancing spatial and computational efficiency, the octree provides an efficient data organization foundation for the entire real-time volume measurement process.

[0103] Dynamic volume calculation: A local coordinate system is constructed, implementing a hybrid convex hull and voxelization strategy. Layered volume calculation is performed using integral projection. A dedicated compensation model for quantifying void fraction using X-ray backscatter data is combined with a dynamic probabilistic error correction model for weighted optimization of multi-scale volume components. The output of the dynamic volume calculation step is a high-precision gravel pile volume value, which has been processed through the convex hull and voxelization hybrid strategy, layered integral projection, void fraction compensation, and multi-scale error correction. This provides accurate foundational data for subsequent intelligent optimization and real-time applications. The goal is to output high-precision, real-time volume values ​​for gravel piles through multi-scale modeling and error correction.

[0104] Intelligent Optimization: Graph Convolutional Networks (GCNs) are trained on sequential point clouds. GCNs are deep learning models specifically designed for processing dynamic point cloud sequences. By combining GCNs with temporal modeling techniques, they capture the dynamic temporal characteristics of 3D point clouds and enable prediction. Gravel deformation trends are predicted using GCNs trained on sequential point clouds. Vibration sensors are used to invert the compaction, and a nonlinear mapping is constructed based on point cloud density to calibrate the final volume.

[0105] Vibration sensors are used to collect the vibration spectrum and amplitude attenuation characteristics of the gravel pile to invert its internal compactness. Combined with the density distribution of the 3D point cloud, a nonlinear mapping model of compactness, density, and volume is constructed to generate a dynamic calibration coefficient. Based on this calibration coefficient, the initial volume calculated using the convex hull-voxelization strategy is adjusted in real time to compensate for porosity deviations caused by loose or dense gravel. For example, the volume estimate is reduced when compactness is high and increased when loose. The calibration results are then verified using X-ray backscatter data, forming a closed "perception-correction-verification" loop. Ultimately, a high-precision volume value optimized through multi-physics collaboration is output, effectively reducing measurement errors.

[0106] The generative adversarial network adopts normal vector constraints. Its discriminator network contains an alternating structure of three-dimensional convolution kernel layers and maximum pooling layers. The generator network contains a U-Net architecture with skip connections. The loss function combines the L1 regularization term and the normal vector direction cosine value. Figure 2 The basic framework of GA-Net is shown in Figure 2, where the discriminator is only used during training. After training, only the generator is retained, with the input being the missing point cloud data and the output being the completed point cloud data. The generator uses the U-Net architecture. Several key formulas are used in training this network. The generator's loss function is:

[0107] .

[0108] .

[0109] .

[0110] ).

[0111] in, represents the number of training samples, that is, the total number of input samples involved in training; To represent the generator network input sample The output result; To indicate the input samples; To indicate the The true label or target output of each sample, that is, the corresponding real 3D point cloud data; express Norm (Manhattan distance), used to measure the difference between the generated results and the true results; Represents the set of normal vectors (or some kind of normal vector error) of all points in the point cloud generated by the generator; Represents the set of normal vectors of all points in the real point cloud; Represents the overall norm of the normal vector in the generated point cloud; is the global norm of the normal vector in the real point cloud, used as a normalization reference.

[0112] Normal vector continuity judgment, the expression corresponding to the normal vector constraint is:

[0113] .

[0114] The curvature clustering segmentation algorithm uses the improved Euclidean clustering method to calculate the point cloud curvature characteristics through principal component analysis and set the adaptive segmentation threshold C threshold =0.5×(σ_max-σ_min) / σ_avg, where σ_max, σ_min, and σ_avg are the maximum, minimum, and average values ​​of the global point cloud curvature, respectively.

[0115] Point cloud neighborhood feature calculation: For any target point , determine its local neighborhood through K nearest neighbor search , construct the covariance moment :

[0116] .

[0117] in is the mean value of the neighborhood point coordinates (i.e. ), the covariance matrix is ​​obtained by eigenvalue decomposition The eigenvalue of ≥ ≥ The eigenvalues ​​represent the distribution intensity of the neighborhood point cloud in three orthogonal directions. is the first eigenvalue; is the second eigenvalue; is the third eigenvalue; Corresponding to the distribution intensity in the main direction, Reflects the degree of discreteness (distribution strength) of the normal vector direction. Reflects the distribution intensity in the subnormal direction.

[0118] in, represents the number of neighborhood points; For the target points; is transposed.

[0119] Point cloud curvature feature calculation:

[0120] .

[0121] Region Growing Clustering: Initializing the Unprocessed Point Set , repeat the following steps until Empty:

[0122] Select The smallest point is used as the seed point to create a new cluster .

[0123] In the neighborhood of the seed point, filter And Euclidean distance Join .

[0124] Iteratively expand the neighborhood of newly added points, dynamically adjusting the search radius:

[0125] .

[0126] When there are no new points in three consecutive iterations, the current cluster growth is completed.

[0127] Post-processing optimization: Eliminate points less than Invalid clustering is performed, quadratic curvature check is performed on the boundary points, and the segmentation results are finally output.

[0128] The construction of the octree spatial topology index maintains a dynamic association with the consistency regions and adopts a non-uniform segmentation strategy based on the consistency region boundaries: when a node is completely contained in a single consistency region, the segmentation in that direction is terminated to maintain the integrity of the regional structure; when a node spans multiple consistency regions and the point cloud density is greater than 100 points / cm³, adaptive segmentation is performed along the region boundary plane with a segmentation depth of D_max=6; bidirectional pointer connections are established between nodes corresponding to adjacent consistency regions, and the pointers contain region attribute similarity tags to accelerate cross-region topology queries.

[0129] like Figure 3 As shown, in the convex hull-voxelization hybrid strategy, the voxel size is dynamically adjusted according to the local curvature, and the adjustment coefficient is set , where k1 is the local surface Gaussian curvature. When k1>0, 0.5cm³ voxels are used, and when k1<0, 1cm³ voxels are used. Based on V (V is the voxel representation of the 3D point cloud model or its reconstructed representation, which refers to the 3D volume information represented by the voxel space or point cloud), an adaptive voxel grid is generated. The ray casting method is used to determine whether the voxel is located inside the model. Then, the number of valid voxels N in each partition is counted, and the partition volume contribution is calculated: ; Finally, add up the volumes of all partitions to get the total volume: .

[0130] The dynamic probabilistic error correction model establishes a triple weighting mechanism: a global convex hull volume weight W_g = 0.3 - 0.1 × (V_proj / V_convex), a local voxel volume weight W_l = 0.4 + 0.2ρ, and a layered projection volume weight W_p = 0.3 + 0.1Δh, where ρ is the local porosity and Δh is the height difference between adjacent projection layers. Three reference volumes are calculated: the global convex hull volume V_convex is obtained using a convex hull algorithm, the volume V_voxel is calculated based on local voxel packing, and the layered projection volume V_proj is obtained through inter-layer projection integration. Normalize the triple weights: calculate the total weight W_total=W_g+W_l+W_p, and obtain the normalized weights W'_g=W_g / W_total, W'_l=W_l / W_total, W'_p=W_p / W_total; finally, calculate the final corrected volume by weighted fusion: calculate the final corrected volume by weighted fusion: V_final=W'_g×V_convex+W'_l×V_voxel+W'_p×V_proj.

[0131] In the field of bulk material measurement, accurate and real-time calculation of gravel pile volume has always been a technical challenge plaguing the industry. Traditional 3D scanning technology, limited by the irregularities and dynamic deformation characteristics of gravel surfaces, often struggles to achieve both accuracy and efficiency. The core breakthrough of this application lies in the construction of a 3D point cloud processing paradigm that integrates multimodal perception and intelligent compensation. Through innovative algorithmic architecture, this approach enables real-time, precise measurement under complex working conditions.

[0132] The physical properties of gravel piles present unique challenges for digital modeling: the random distribution of surface particles results in multiple reflections and diffuse reflections during laser scanning, and the point cloud data is interspersed with a large amount of noise and missing areas. Conventional denoising algorithms, while removing discrete noise, often destroy the microscopic geometric features of the gravel pile. To address this, this application proposes a generative adversarial network (GA-Net) based on normal vector constraints, which innovatively uses the continuity of the point cloud normal vector field as a physical constraint for the generator. This network extracts multi-scale spatial features through three-dimensional convolutional kernel layers and utilizes a U-Net architecture with skip connections to preserve detailed information. During adversarial training, the normal vector direction of the repaired area is forced to remain topologically consistent with the surrounding point cloud. Experiments show that compared to traditional Poisson reconstruction methods, this technology improves the repair accuracy of dust-occluded areas by 42% and effectively preserves tiny raised features.

[0133] Faced with the complex multi-level void structure inside the gravel pile, traditional volume calculation methods often fall into modeling difficulties. Fixed-resolution voxelization methods either underestimate the void volume due to insufficient accuracy, or lose real-time performance due to excessive computational complexity. The dynamic volume calculation module of this application creatively proposes a convex hull-voxelization hybrid strategy: first, a minimum convex hull is constructed in the local coordinate system to capture the macroscopic contour, and then the voxel resolution is dynamically adjusted according to the curvature characteristics - high-precision voxels are used to capture microscopic voids in areas where the curvature changes drastically, and medium-precision voxels are switched in flat areas to improve computational efficiency. This divide-and-conquer strategy of "macroscopic delimitation and microscopic refinement" controls the error in void volume calculation to below 2%.

[0134] To enable engineering applications in industrial sites, the entire algorithm architecture has been deeply optimized. During point cloud preprocessing, an improved curvature clustering segmentation algorithm extracts local geometric features through principal component analysis and sets adaptive thresholds to accurately segment sand and gravel sub-regions. The octree spatial index employs a non-uniform segmentation strategy, increasing spatial query efficiency by more than three times while maintaining topological connectivity. The dynamic probabilistic error correction model innovatively introduces a triple weight allocation mechanism. By analyzing the statistical correlation between the global convex hull volume, local voxel volume, and layered projection volume, a Bayesian optimization framework is constructed to enable self-verification and correction of measurement results.

[0135] After rigorous field validation, this application demonstrated significant performance advantages: in tests with gravel piles containing 5mm-50mm particles, it maintained stable operation in industrial environments ranging from -20°C to 55°C. The 3D visualization interface also displayed the volume change curve of the gravel pile in real time.

[0136] This application not only addresses the industry pain point of bulk material volume measurement, but also provides a universal solution for 3D perception in complex scenarios through its innovative algorithmic framework. Through the deep cross-integration of multidisciplinary technologies, this application successfully bridges the gap between theoretical models and engineering applications, laying a key technological foundation for the digital transformation of areas such as smart construction sites.

[0137] The specific implementation process of this application can be divided into four core stages according to the engineering implementation process: hardware system deployment, dynamic point cloud acquisition, intelligent point cloud processing, and real-time volume calculation. The following is a detailed explanation of the technical solution based on industrial field implementation cases.

[0138] When implementing on-site in an industrial environment, the first step is to build a track-type 3D scanning hardware system.

[0139] The hardware system includes:

[0140] Track scanning device: includes lidar, X-ray backscatter sensor, and IMU inertial measurement unit.

[0141] Preprocessing module: equipped with hardware-accelerated GA-Net processor and curvature clustering dedicated computing unit.

[0142] Dynamic computing core: integrated convex hull fast generator and programmable voxelization engine.

[0143] Human-computer interaction terminal: equipped with real-time three-dimensional visualization interface.

[0144] Double-track running tracks are laid parallel to both sides of the gravel stockpile. The track spacing is determined by the maximum stockpile width (typically 8-12 meters). The track trolley integrates a 32-line LiDAR, an IMU inertial navigation module, and an edge computing unit. The LiDAR is mounted on a tiltable gimbal, with a scanning angle set to ±30° in pitch to accommodate varying stockpile heights. Upon power-up, the system performs a self-test: the LiDAR is calibrated using a pre-set calibration plate, and the IMU module performs a heading zero calibration. During the debugging phase, the matching relationship between the trolley's travel speed and scanning frequency must be verified to ensure a point cloud overlap of at least 70% at a speed of 0.5 m / s.

[0145] During dynamic scanning, the trolley moves back and forth along the track at a constant speed. The LiDAR emits a scanning beam at a frequency of 10 Hz, and each frame of the point cloud contains approximately 800,000 raw points. The point cloud preprocessing module performs coordinate transformation in real time, converting the point cloud data from the LiDAR coordinate system to the global coordinate system. To address the multipath interference caused by particles on the surface of the gravel pile, a dual-threshold filtering algorithm is designed: first, distance mutation points (distance between adjacent points exceeding three standard deviations) are filtered out, and then an improved radius filter is used to retain points with a density above the threshold within a 0.1m neighborhood. The preprocessed point cloud is transmitted to the edge computing unit via a 5G wireless module, and the timestamp and pose information are stored.

[0146] During the point cloud inpainting phase, a GA-Net generative adversarial network (GA-Net) was used to address dust-occluded areas. During network training, a database of 100,000 sand and gravel point cloud samples was constructed, each consisting of a complete point cloud and artificially simulated occlusion data. The generator uses a U-Net architecture, taking as input 256×256×256 voxelized local point cloud blocks. Multi-scale features are extracted through 3D convolutional layers. The discriminator is designed as a two-channel structure, simultaneously evaluating the geometric consistency of the point cloud and the continuity of the normal field. The loss function consists of an L2 reconstruction loss, a normal cosine similarity loss, and an adversarial loss, with a weighting ratio of 3:2:1. During the deployment phase, voids in the point cloud are detected in real time. GA-Net is directly used to inpaint missing areas with a diameter less than 0.5m. Larger voids are first processed by boundary extraction and then by segmentation. The inpainted point cloud undergoes physical plausibility verification, checking whether the surface curvature abrupt change points are within the statistical distribution of sand and gravel particle sizes.

[0147] The volume calculation module uses a hybrid convex hull and voxelization strategy. First, the minimum convex hull of the gravel pile is constructed in the global coordinate system, and the initial volume V_convex is generated using the QuickHull algorithm. The internal space of the convex hull is then discretized into a dynamic voxel grid with a base resolution of 5 cm³. During the curvature feature extraction phase, the eigenvalues ​​of the covariance matrix are calculated for each point. Regions with a curvature value λ3 / (λ1+λ2+λ3) greater than 0.15 are marked as high curvature regions, and the voxel resolution in these regions is automatically increased to 1 cm³. The eigenvalues ​​λ1, λ2, and λ3 describe the distribution intensity or degree of dispersion of neighborhood points in three orthogonal directions. λ1 is the maximum eigenvalue, indicating the degree of dispersion of the point cloud in the most principal direction of the neighborhood; λ2 is the intermediate eigenvalue, indicating the degree of dispersion in the second principal direction; and λ3 is the minimum eigenvalue, indicating the degree of dispersion of neighborhood points perpendicular to the local principal plane. Voxel occupancy is determined using a ray casting method: a probe ray is fired from the voxel center in six positive directions, and the number of intersections with the point cloud surface is counted. Any voxel with an odd number of intersections is considered internal. The final volume V_real = V_convex × α + V_voxel × β, where α = 0.22 and β = 0.78 are compensation coefficients calibrated through extensive experiments.

[0148] To achieve engineering-grade real-time performance, the algorithm architecture undergoes multi-level optimization. Point cloud segmentation utilizes an improved curvature clustering algorithm: For each point, the standard deviation of the normal vector within its local neighborhood is calculated. Points exceeding a threshold are identified as feature edges. Connected regions are merged using a region-growing algorithm, ultimately segmenting the point cloud into several sub-heaps. The octree spatial index employs a non-uniform construction strategy, partitioning dense areas of the point cloud up to the eighth level (1cm resolution) and sparse areas only to the fifth level (16cm resolution). A dynamic probabilistic error correction model analyzes historical measurement data in real time to establish a three-level confidence assessment for volume calculations: a prior probability P1 is calculated based on the statistical distribution of the convex hull volume; a likelihood probability P2 is calculated based on the current voxel fill rate; and a posterior probability P3 is calculated based on the layered projected area. A Bayesian optimizer integrates these three factors using a weight matrix W = [0.4, 0.3, 0.3]. Recalculation is triggered when the combined confidence falls below 85%.

[0149] After system deployment, on-site calibration and verification are required. Long-term operational testing has shown that, in an environment with a dust concentration below 50 mg / m³, the system can operate continuously for over eight hours, maintaining a stable volume refresh rate of 1 Hz. After six months of practical engineering application, this technical solution has reduced volume calculation errors by 76% compared to traditional laser scanning methods, extending the equipment failure interval to over 1,200 hours, and significantly improving the intelligent management of bulk materials.

[0150] Based on the same inventive concept, embodiments of the present application also provide a device for determining the volume of a three-dimensional point cloud of a gravel pile based on a railcar, for implementing the aforementioned method for determining the volume of a three-dimensional point cloud of a gravel pile based on a railcar. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for determining the volume of a three-dimensional point cloud of a gravel pile based on a railcar provided below can be found in the aforementioned limitations of the method for determining the volume of a three-dimensional point cloud of a gravel pile based on a railcar, and will not be repeated here.

[0151] In an exemplary embodiment, a device for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley is provided, comprising:

[0152] The data acquisition module is used to obtain the original point cloud data of the gravel pile. The original point cloud data is obtained by scanning and collecting using a track trolley.

[0153] The processing module is used to preprocess the original point cloud data and use the curvature clustering segmentation algorithm to divide the preprocessed point cloud data into regions to obtain the divided point cloud data.

[0154] The repair module is used to use an adversarial network to perform regional repair on the divided point cloud data to obtain repaired point cloud data; the adversarial network is determined based on the normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field.

[0155] The partitioning module is used to spatially partition the repaired point cloud data based on the octree spatial topology index to obtain point cloud data at different levels, and determine the topological connection relationship between different levels to obtain point cloud subsets.

[0156] The calculation module is used to adopt a convex hull-voxelization hybrid strategy to perform integral projection layered calculation based on point cloud subsets to obtain the initial volume of the gravel pile.

[0157] The calibration module is used to collaboratively optimize and calibrate the initial volume of the gravel pile using a time-series point cloud training graph convolutional network to obtain a calibrated volume. The time-series point cloud training graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile. After performing internal compactness inversion, it combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness-density-volume.

[0158] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, a memory, an input / output (I / O) interface, and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The I / O interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When executed by the processor, the computer program implements a method for determining the volume of a three-dimensional point cloud of a gravel pile.

[0159] Those skilled in the art will understand that the structure shown above is only a partial structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown above, or combine certain components, or have a different component arrangement.

[0160] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0161] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0162] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0163] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0164] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0165] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0166] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0167] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley, characterized in that: include: Obtain the original point cloud data of the gravel pile; The original point cloud data is obtained by scanning and collecting based on the track trolley; Preprocessing the original point cloud data, and dividing the preprocessed point cloud data into regions using a curvature clustering segmentation algorithm to obtain divided point cloud data; An adversarial network is used to perform regional repair on the divided point cloud data to obtain repaired point cloud data; the adversarial network is determined based on a normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field; Performing spatial division on the repaired point cloud data based on an octree spatial topology index to obtain point cloud data of different levels, and determining topological connection relationships between different levels to obtain point cloud subsets; A convex hull-voxelization hybrid strategy is used to perform integral projection layered calculation based on the point cloud subset to obtain the initial volume of the gravel pile. A time-series point cloud training graph convolutional network is used to collaboratively optimize and calibrate the initial volume of the gravel pile to obtain a calibrated volume. The time-series point cloud training graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile. After performing internal compactness inversion, the network combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness-density-volume. The adversarial network includes: a discriminator and a generator; the discriminator includes an alternating structure of a three-dimensional convolution kernel layer and a maximum pooling layer; the generator includes a U-Net architecture with skip connections; The loss function of the generator is determined based on the L1 regularization term and the normal vector direction cosine value using the normal vector constraint; The expression of the loss function is: ； The expression corresponding to the normal vector constraint is: ； in, is the loss function; To rebuild losses; is the direction cosine value of the normal vector; To combat losses; is the normal vector constraint; is the expectation of the distribution of real point cloud data; It is the subnetwork in the discriminator that processes the normal vector field and is used to evaluate the authenticity of the normal vector field; is the normal vector field corresponding to the real point cloud; is the expectation of the noise distribution of the generator input; A fake point cloud generated by the generator based on noise; Fake point cloud The corresponding normal vector field; A convex hull-voxelization hybrid strategy is used to perform integral projection layered calculation based on the point cloud subset to obtain the initial volume of the gravel pile, specifically including: Based on the Gaussian curvature of the local surface corresponding to the point cloud subset, according to Adjust the voxel size; where α is the voxel size coefficient; k1 is the local surface Gaussian curvature; determining a voxel grid according to the adjusted voxel size; Using a ray casting method, a projection layering process is performed based on the voxel grid to obtain a plurality of partitions; According to the number of voxels in all partitions, the convex hull algorithm is used to perform integral accumulation processing to obtain the initial volume of the gravel pile.

2. The method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley according to claim 1, characterized in that: The original point cloud data is preprocessed, and the preprocessed point cloud data is divided into regions using a curvature clustering segmentation algorithm to obtain divided point cloud data, specifically including: Performing multi-scale noise filtering on the original point cloud data to obtain preprocessed point cloud data; The curvature clustering segmentation algorithm is used to divide the preprocessed point cloud data into regions to obtain the divided point cloud data.

3. The method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley according to claim 2, characterized in that: The curvature clustering segmentation algorithm is used to divide the pre-processed point cloud data into regions, and the divided point cloud data is obtained, including: An improved Euclidean clustering method is used to calculate the point cloud curvature feature through principal component analysis; the point cloud curvature feature is determined based on the point cloud neighborhood feature; The method for determining the point cloud curvature feature specifically includes: For any target point in the preprocessed point cloud data, the local neighborhood is determined by K-nearest neighbor search and the covariance moment is constructed; Based on the covariance moment, the eigenvalues ​​of the covariance matrix are obtained by eigenvalue decomposition; the eigenvalues ​​represent the distribution intensity of the neighborhood point cloud in three orthogonal directions; Based on the eigenvalues, the formula is adopted: ； Determine the curvature characteristics of the point cloud; For all target points in the preprocessed point cloud data, select the target point with the smallest value of the point cloud curvature feature as the seed point and create a new cluster ; In the neighborhood of the seed point, filter and Join ; right The neighborhood is expanded iteratively for the newly added points in and the search radius is adjusted dynamically: ； When there are no new points after a set number of consecutive iterations, cluster growth is completed; Eliminate seed points when the number of points is less than Clustering, perform quadratic curvature check on the cluster boundary points to obtain the divided point cloud data; in, For the The value of the point cloud curvature feature corresponding to the target point; is the first eigenvalue; is the second eigenvalue; is the third eigenvalue; For the The value of the point cloud curvature feature corresponding to the target point; To set the adaptive segmentation threshold; For the The target point and The Euclidean distance between target points; is the search radius; The lower limit of the number of seed points.

4. The method for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley according to claim 1, characterized in that: The temporal point cloud training graph convolutional network adopts a triple weight distribution mechanism to perform weighted fusion processing on the initial volume of the gravel pile to achieve collaborative optimization calibration and obtain a calibrated volume; wherein the triple weight includes: global convex hull volume weight, local voxel volume weight and layered projection volume weight.

5. A device for determining the volume of a three-dimensional point cloud of a gravel pile based on a rail trolley, characterized in that: include: A data acquisition module is used to obtain the original point cloud data of the sand and gravel pile; The original point cloud data is obtained by scanning and collecting based on the track trolley; a processing module, configured to pre-process the original point cloud data and perform region division on the pre-processed point cloud data using a curvature clustering segmentation algorithm to obtain divided point cloud data; A restoration module is configured to perform regional restoration on the divided point cloud data using an adversarial network to obtain restored point cloud data; the adversarial network is determined based on a normal vector constraint; the normal vector constraint is determined based on the continuity of the point cloud normal vector field; a partitioning module, configured to spatially partition the repaired point cloud data based on an octree spatial topology index to obtain point cloud data at different levels, and determine the topological connection relationship between the different levels to obtain point cloud subsets; a calculation module, configured to adopt a convex hull-voxelization hybrid strategy to perform integral projection layered calculation based on the point cloud subset to obtain an initial volume of the gravel pile; A calibration module is configured to collaboratively optimize and calibrate the initial volume of the gravel pile using a time-series point cloud training graph convolutional network to obtain a calibrated volume; the time-series point cloud training graph convolutional network is based on the vibration spectrum and amplitude attenuation characteristics of the gravel pile, performs internal compactness inversion, and then combines the density distribution corresponding to the three-dimensional point cloud data to construct a nonlinear mapping model of compactness-density-volume; The adversarial network includes: a discriminator and a generator; the discriminator includes an alternating structure of three-dimensional convolution kernel layers and maximum pooling layers; the generator includes a U-Net architecture with skip connections; The generator's loss function is determined based on the normal vector constraint and the L1 regularization term and the normal vector direction cosine value; The expression of the loss function is: ； The expression corresponding to the normal vector constraint is: ； in, is the loss function; To rebuild losses; is the direction cosine value of the normal vector; To combat losses; is the normal vector constraint; is the expectation of the distribution of real point cloud data; It is the subnetwork in the discriminator that processes the normal vector field and is used to evaluate the authenticity of the normal vector field; is the normal vector field corresponding to the real point cloud; is the expectation of the noise distribution of the generator input; A fake point cloud generated by the generator based on noise; Fake point cloud The corresponding normal vector field; A convex hull-voxelization hybrid strategy is used to perform integral projection layered calculations based on point cloud subsets to obtain the initial volume of the gravel pile, including: Based on the Gaussian curvature of the local surface corresponding to the point cloud subset, according to Adjust the voxel size; where α is the voxel size coefficient; k1 is the local surface Gaussian curvature; The voxel grid is determined according to the adjusted voxel size. The ray casting method is used to perform projection and layering processing based on the voxel grid to obtain multiple partitions. The convex hull algorithm is used to perform integral accumulation processing based on the number of voxels in all partitions to obtain the initial volume of the gravel pile.

6. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for determining the three-dimensional point cloud volume of a gravel pile based on a rail trolley according to any one of claims 1 to 4.

Images